多姿态人脸识别的改进PCA方法

An Adapted PCA to Multiview Face Recognition

本文研究了在每个对象只有一幅图像可用的情况下进行人脸识别。由于姿态变化所造成的自身遮掩和旋转的非线性,人脸识别的准确率将大大降低,广泛采用的主分量分析方法性能也将随之下降。通过分析用主分量分析生成的正面特征空间中的权值变化,我们发现给这些权值矢量加上系数,系统性能将大大提高。从而,我们提出了多姿态主分量分析方法。在UMIST人脸库上的实验结果表明我们的方法优于著名的特征脸方法,它大大的提高了识别率。

This paper conducts face recognition task when only one frontal face image for each person is available. While pose variation can greatly reduce recognition accuracy due to occlusion and rotation nonlinearity, the widely used method principle component analysis （PCA） is subject to serious reduction in performance. With analyzing the weights in the frontal eigenspace generated by PCA, the authors found that the performance can he largely enhanced hy assigning a pose factor to the PCA weight vector. Consequently a new method called muhiview PCA is proposed. Experimental results on UMIST database show the advantages of this method over traditional eigenface method, it can improve recognition accuracy greatly.